Sync from GitHub · 2026-08-29 08:44 UTC

routes/chat.py

@@ -69,8 +69,8 @@ def chat_message(
 ):
     """
     Accepts a user message and returns a streaming SSE response.
-    For passive AI Chat: uses light model (llama-3.1-8b-instant).
-    For AI Consultation: uses heavy model (llama-3.3-70b-versatile).
+    For passive AI Chat: uses light model (LIGHT_MODEL).
+    For AI Consultation: uses heavy model (HEAVY_MODEL).
 
     Stream format (SSE):
       data: {"type": "thinking", "token": "..."}   ← reasoning (heavy model only)

routes/transcribe.py

@@ -1,10 +1,8 @@
 from fastapi import APIRouter, UploadFile, File, Depends, HTTPException
 from auth import verify_firebase_token
-import os
-from groq import Groq
+from utils.groq_client import call_groq_transcribe
 
 router = APIRouter()
-groq_client = Groq(api_key=os.environ.get("GROQ_API_KEY"))
 
 SUPPORTED_AUDIO_TYPES = {
     "audio/webm", "audio/mp4", "audio/mpeg", "audio/wav",
@@ -49,8 +47,7 @@ async def transcribe_audio(
         raise HTTPException(status_code=400, detail="Audio file is empty.")
 
     try:
-        transcription = groq_client.audio.transcriptions.create(
-            model="whisper-large-v3-turbo",
+        transcription = call_groq_transcribe(
             file=(audio.filename or "recording.webm", audio_bytes, content_type),
             response_format="json",
             language="en",

utils/groq_client.py

@@ -39,8 +39,11 @@ if not _KEYS:
 print(f"[groq] {len(_KEYS)} API key(s) loaded for rotation")
 
 # ─── Model names ──────────────────────────────────────────────────────────────
-HEAVY_MODEL = "llama-3.3-70b-versatile"   # reasoning, profile generation, consultation
-LIGHT_MODEL  = "llama-3.1-8b-instant"     # chat, quick responses
+# llama-3.3-70b-versatile / llama-3.1-8b-instant were retired by Groq — replaced
+# with the current openai/gpt-oss lineup (see Groq's models page for the live list).
+HEAVY_MODEL = "openai/gpt-oss-120b"   # reasoning, profile generation, consultation
+LIGHT_MODEL  = "openai/gpt-oss-20b"   # chat, quick responses
+TRANSCRIBE_MODEL = "whisper-large-v3-turbo"
 
 # ─── Round-robin state ────────────────────────────────────────────────────────
 _idx  = 0
@@ -121,6 +124,47 @@ def call_groq(
     raise last_error or RuntimeError("All Groq API keys are rate-limited")
 
 
+def call_groq_transcribe(
+    file: tuple,
+    model: str = TRANSCRIBE_MODEL,
+    response_format: str = "json",
+    language: str | None = None,
+):
+    """
+    Calls Groq's Whisper transcription endpoint with the same key-rotation
+    on 429 errors as the chat completion calls above.
+
+    `file` is the (filename, bytes, content_type) tuple the Groq SDK expects.
+    """
+    last_error: Exception | None = None
+
+    for attempt in range(len(_KEYS) + 1):
+        try:
+            client = Groq(api_key=_current_key())
+            kwargs: dict = dict(model=model, file=file, response_format=response_format)
+            if language:
+                kwargs["language"] = language
+            return client.audio.transcriptions.create(**kwargs)
+
+        except Exception as e:
+            last_error = e
+            err_str = str(e)
+            is_rate_limit = (
+                '429' in err_str
+                or 'rate_limit' in err_str.lower()
+                or 'rate limit' in err_str.lower()
+            )
+
+            if is_rate_limit and attempt < len(_KEYS):
+                _rotate()
+                time.sleep(1)
+                continue
+
+            raise
+
+    raise last_error or RuntimeError("All Groq API keys are rate-limited")
+
+
 def call_groq_json(
     system_prompt: str,
     user_prompt: str,